CS5112: Algorithms and Data Structures for Applications

Transcription

1 CS5112: Algorithms and Data Structures for Applications Lecture 14: Exponential decay; convolution Ramin Zabih Some content from: Piotr Indyk; Wikipedia/Google image search; J. Leskovec, A. Rajaraman, J. Ullman: Mining of Massive Datasets,

2 Administrivia Q7 delayed due to Columbus day holiday HW3 out but short, can do HW in groups of 3 from now on Class 10/23 will be prelim review for 10/25 Greg lectures on dynamic programming next week Anonymous survey out, please respond! Automatic grading apology

3 Survey feedback

4 Automatic grading apology In general students should not have a grade revised downward Main exception is regrade requests Automatic grading means that this sometimes happened At the end, we decided that the priority was to assign HW grades based on how correct the code was Going forward, please treat HW grades as tentative grades We will announce when those grades are finalized After, they will only be changed under exceptional circumstances

5 Today Two streaming algorithms Convolution

6 Some natural histogram queries Top-k most frequent elements Find all elements with frequency > 0.1% What is the frequency of element 3? What is the total frequency of elements between 8 and 14? How many elements have non-zero frequency?

7 Streaming algorithms (recap) 1. Boyer-Moore majority algorithm 2. Misra-Gries frequent items 3. Find popular recent items (box filter) 4. Find popular recent items (exponential window) 5. Flajolet-Martin number of items

8 Weighted average in a sliding window Computing the average of the last k inputs can be viewed as a dot product with a constant vector v = 1, 1,, 1 k k k Sometimes called a box filter Easy to visualize This is also a natural way to smooth, e.g., a histogram To average together adjacent bins, v = [ 1 3, 1 3, 1 3 ] This kind of weighted average has a famous name: convolution

9 Decaying windows Let our input at time t be a 1, a 2,, a t With a box filter over all of these elements we computed t i=0 Instead let us pick a small constant c and compute 1 t 1 t 1 a t i a t = a t i 1 c i i=0

10 Easy to update this Update rule is simple, let the current dot product be a t a t+1 = 1 c a t + a t+1 This downscales the previous elements correctly, and the new element is scaled by 1 c 0 = 1 This avoids falling off the edge Gives us an easy way to find popular items

11 4. Popular items with decaying windows We keep a small number of weighted sum counters When a new item arrives for which we already have a counter, we update it using decaying windows, and update all counters How do we avoid getting an unbounded number of counters? We set a threshold, say ½, and if any counter goes below that value we throw it away The number of counters is bounded by 2 c

12 How many distinct items are there? This tells you the size of the histogram, among other things To solve this problem exactly requires space that is linear in the size of the input stream Impractical for many applications Instead we will compute an efficient estimate via hashing

13 5. Flajolet-Martin algorithm Basic idea: the more different elements we see, the more different hash values we will see We will pick a hash function that spreads out the input elements Typically uses universal hashing

14 Flajolet-Martin algorithm Pick a hash function h that maps each of the n elements to at least log 2 n bits For input a, let r(a) be the number of trailing 0s in h(a) r(a) = position of first 1 counting from the right E.g., say h(a) = 12, then 12 is 1100 in binary, so r(a) = 2 Record R = the maximum r(a) seen Estimated number of distinct elements = 2 R Anyone see the problem here? J. Leskovec, A. Rajaraman, J. Ullman: Mining of Massive Datasets, 14

15 Why It Works: Intuition Very rough intuition why Flajolet-Martin works: h(a) hashes a with equal probability to any of n values Sequence of (log 2 n) bits; 2 r fraction of a s have tail of r zeros About 50% hash to ***0 About 25% hash to **00 So, if we saw the longest tail of r=2 (i.e., item hash ending *100) then we have probably seen about 4 distinct items so far Hash about 2 r items before we see one with zero-suffix of length r J. Leskovec, A. Rajaraman, J. Ullman: Mining of Massive Datasets, 15

16 Convolution Weighted average with a mask/stencil/template Dot product of vectors Many important properties and applications Symmetric in the inputs Equivalent to linear shift-invariant systems Well behaved, in a certain precise sense Primary uses are smoothing and matching This is the C in CNN

17 Local averaging in action

18 Smoothing parameter effects

19 Matched filters Convolution can be used to find pulses This is actually closely related to smoothing How do we find a known pulse in a signal? Convolve the signal with our template! E.g. to find something in the signal that looks like [1 6-10] we convolve with [1 6-10] Question: what sense does this make? Anecdotally it worked for finding boxes

20 Box finding example

21 Pulse finding example

22 Why does this work? Some nice optimality properties, but the way I described it, the algorithm fails Idea: the [1 6-10] template gives biggest response when signal is [ ] Value is 137 at this point But is this actually correct? You actually need both the template and the input to have a zero mean and unit energy (sum of squares) Easily accomplished: subtract -1, then divide by 137, get 1/137 * [2 7-9]

23 Geometric intuition Taking the dot product of two vectors Recall a b c [e f g] = ae + bf + cg Basically the projection of a vector on another The normalized vector with the biggest projection on x is, of course: x! 23